Database In-Memory in 23ai

The 10-Year Anniversary Deep Dive

Shasank Chavan – VP of In-Memory, Data and AI Technologies

Ambili Thottathil – Principal Consultant - Analytics, Rittman Mead
Tirthankar Lahiri – SVP of Mission-Critical Data and AI Engines

1 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory
Released July, 2014 - 10 years ago

"We're looking
forward to a future
of real-time
analytics at ungodly
speeds."
2 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates
 Real-Time Enterprises need In-Memory

"The big data revolution is the

ability to make decisions using
all of your data, not just some
of your data. We're looking
forward to a future of real-
time analytics at ungodly
speeds.”
- Larry Ellison

3 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory
In a nutshell

4 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Analytic and OLTP Formats
What’s your favorite data format?

SALES
Row Format
Fast for OLTP!
Slower for Analytics

Dual Format
(+) • Best for both Fast Analytics
and fast OLTP (No need for
SALES Analytic indexes)

Column Format
Fast for Analytics!
REALLY slow for OLTP!

5 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Breakthrough: Dual Format Database

Buffer Cache
In-Memory • BOTH row and column formats
Column Store
for same table
• Simultaneously active and
transactionally consistent
Benefits Benefits analytics
OLTP and reporting • Analytics & reporting use new
in-memory Column format

• OLTP uses proven row format

• No application changes
Operational
Data required

6 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory Technology
Scanning and filtering data more efficiently

Columnar Format Compression Storage Indexes SIMD Vector

Processing
Example:
Find sales in
State of California

STATE
CPU
CA

Vector Register
Load Vector
CA
multiple CA
Compare
State all values
values in 1 cycle
CA

> 100x Faster

Access only the Scan & filter data in Prune out any Process multiple
columns you need compressed format unnecessary data column values in a
from the column single CPU instruction

7 Oracle CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Accelerate Analytics
Improves all aspects of analytic queries

Data Scans Joins In-Memory

Aggregation
STORE_ID

AMOUNT

Speed of memory Convert Star Joins into Create In-Memory

Scan and Filter only the 10X Faster Column Scans Report Outline that is
needed Columns Search large table for Populated during Fast Scan

Vector Instructions values that match small Runs Reports Instantly

table

8 Oracle CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Accelerate Mixed Workloads

1–3 10 – 20 1–3
OLTP Analytic OLTP
Indexes Indexes Indexes
In-Memory
Column Store
Table Table

REPLACE

• Inserting one row into a table requires • Column Store not persistent so updates are: Fast!
updating 10-20 analytic indexes: Slow!
• Fast analytics on any columns
• Fast analytics only on indexed columns
• No analytic indexes: Reduces database size
• Analytic indexes increase database size

9 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory
Top-5 Innovations

10 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory | Dual-Format Architecture
#1

In-Memory
Buffer Cache
“When you update one Column Store

you always update the

other…the data is Benefits Benefits analytics
OLTP and reporting

consistent between
those two formats.”
- Larry Ellison
Operational
Data

11 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Mixed Workloads
Dual-Format Architecture enables fast Mixed Workloads and faster Memory Memory
Analytics
X .… X X
Fast In-Memory DML because invalid row is logically removed from
SALES SALES
column store (just set a bit)
Row Column
Format Format
Analytic query will ignore invalid rows in column store, and just vector
process valid rows.

Invalid rows are then processed.

SALES

Mixed workload performance can suffer if the number of invalid rows

accumulates in IMCUs
• Additional techniques to refresh a dirty IMCU in the background

12 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory | Performance
#2
In-Memory Vectorized Joins

ITEMS
“When you put data in
memory, one of the Example:
Join Sales and Items
Vector Register
reasons you do that is SALES

Vector Register
to make it go faster.”
- Larry Ellison

13 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory | Automation
#3
Automatic In-Memory Management

In-Memory Column Store

“Flip a switch and all of your Hot Data Intermediate Data Cool Data

applications run much, much

faster. Every application you
wrote, every application you
Remove
Cold Table

bought, runs without a single

React

change.”
Populate
Remove cold data,
Hot Table
populate hot data

- Larry Ellison

Hot Data

14 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory | Exadata Cell-Memory
#4
Cell-Memory Smart Scans

“It’s the fastest machine In-Memory Columnar scans

Up to 2TB DRAM per Database
Server

in the world for

databases stored in
memory” In-Flash Columnar scans
Up to 25.6TB Flash per Storage
- Larry Ellison Server

15 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory | Converged Workloads
#5
1. Store Spatial 2. Store Optimized Text 3. Store JSON in optimized
Summaries in Column Index structure in Column binary representation in
Store for Faster Filtering Store for fast searches Column Store

In-Memory (IM) In-Memory (IM) {

Table Columns Table Columns "Theater":"AMC 15",
"Movie":”Rogue One",
"Time“:2017-01-09 18:45",
Spatial Spatial Resume Text
"Tickets":{
Column Summary (Text) Index
"Adults":2
}
Words }
..
database
..
…
00101001…

16 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 What’s New in 23ai ?

17 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Oracle 23ai is the sum of…

21 +
c
All the features from Oracle
21c Innovation Release
300+ New Features and
Enhancements

Key focus areas:

JSON, Graph, Vector,
Microservices, Developer
Productivity
18 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates
 New in 21c for Database In-Memory
Base Level Cell Memory Enhanced External In-Memory In-Memory
Feature Only Table Support Vectorized Joins Hybrid Scans

AIM In-Memory In-Memory In-Memory

Fully Automatic JSON Data Type Full Text Columns Spatial

19 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Base Level Feature
Customers can use up to a 16GB IM column
System Global Area (SGA)
store without having to license the Database
In-Memory option. Large In-Memory
Pool
• INMEMORY_FORCE = BASE_LEVEL Area
Buffer (Max 16GB)

The purpose of this Base Level feature is to Cache

allow customers to see the value of Database Other

In-Memory.
Shared
Not all Database In-Memory features are Pool
Log
available with the Base Level feature. Buffer

Available in 21c and starting with the 19.8 RU

20 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory Advisor

21 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 New Database Embedded In-Memory Advisor
• Previous version
• Standalone package that had to be installed in the database
• Analyzed existing database workload using AWR/ASH data to determine benefit that Database In-Memory might provide
• Generated an HTML report with estimates for overall benefit, benefit for top SQL statements and objects based on an in-
memor size

• New in 23ai, the In-Memory Advisor is now a package in Oracle Database

• Relies on Heat Map data for analysis
• Heat Map is now available as part of Oracle Database Enterprise Edition (no separate license required)

• Composed of two tools:

• A new Eligibility tool to quickly identify databases where Database In-Memory would NOT be useful – also available in
19c (19.20 RU)
• A Comprehensive Analysis tool based on a specified workload timeframe that produces an object benefit analysis

22 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory Eligibility Tool – Example 1
Database running TPC-C workload.
Identify representative workload through AWR snapshots ids:
SQL> variable inmemory_eligible BOOLEAN
SQL> variable analysis_summary VARCHAR2(100)
SQL> exec dbms_inmemory_advise.is_inmemory_eligible(1, 100, :inmemory_eligible,
:analysis_summary);
SQL> print inmemory_eligible
INMEMORY_ELIGIBLE
-----------------
FALSE
SQL> print analysis_summary
ANALYSIS_SUMMARY
----------------
Only 5.15% of the workload is analytical

24 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory Eligibility Tool – Example 2
Database running TPC-H workload.
Identify representative workload through Start and End times
SQL> variable inmemory_eligible BOOLEAN
SQL> variable analysis_summary VARCHAR2(100)
SQL> exec dbms_inmemory_advise.is_inmemory_eligible(sysdate - 7, sysdate,
:inmemory_eligible, :analysis_summary);
SQL> print inmemory_eligible
INMEMORY_ELIGIBLE
-----------------
TRUE Backported
SQL> print analysis_summary to 19c !!
ANALYSIS_SUMMARY
----------------
100% of the workload is analytical

25 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Task_ Column DB Recommended
In-Memory Advisor 2.0 – Example Id Store
Size in
Time
in Secs
object list

MB
Step 1. Customer starts advisor tracking
SQL> exec 1 0 6000
dbms_inmemory_advisor.start_tracking(); 1 1000 4500 Table Owner:
TPCH, Table Name:
Step 2. Customer stops advisor tracking LINEORDER,
PARTITION NAME
2023_MAY;
SQL> exec
dbms_inmemory_advisor.stop_tracking();
1 2000 3000 Table Owner:
Step 3. Customer generates report advisor tracking TPCH, Table Name:
LINEORDER,
SQL> exec PARTITION NAME
2023_MAY;
dbms_inmemory_advisor.generate_advise(); Table Owner:
TPCH, Table Name:
Step 4. Customer queries the view LINEORDER,
PARTITION NAME
dba_inmemory_advise_recommendation to view 2023_APR;
recommendation.
SQL> select * from
dba_inmemory_advisor_recommendation;
26 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates
 Automatic In-Memory
Management

27 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Automatic Enablement of In-Memory Features
• In Oracle Database 23ai AIM has been enhanced to add the
ability to automatically:
• Optimized Arithmetic
• Bloom filter optimization (cached hash values) Capture
• Stored using vector optimization
• Join Group

• Requires INMEMORY_AUTOMATIC_LEVEL be set to HIGH

Monitor Identify
• Column store cannot be under memory pressure
• Reduces manual effort required to leverage key
performance features
• Enhanced workload analysis to better account for mixed
workload environments Decide Verify
• DML overhead with In-Memory (for example, fetching invalid rows
in column store from buffer cache) is factored into analysis

• No application changes required

28 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Automatic In-Memory Sizing
Automatic Sizing of the In-Memory Column Store
SGA
• Auto-size column store to factor in different workload types
sharing resources at the same time
• Shrink column store and increase buffer cache if DML In-Memory Buffer
intensive workload predicted.
• Grow column store and shrink buffer cache if OLAP
Column Store Cache
intensive workload predicted
• Works with Automatic In-Memory (AIM)
• Requires AIM level HIGH or MEDIUM OLTP OLAP
• ASMM to be enabled (i.e. SGA_TARGET)
• VECTOR_MEMORY_SIZE not specified or set to 0

• The INMEMORY_SIZE parameter sets the minimum IM

SGA
column store size
In-Memory
Works in conjunction with In-Memory Hybrid Exadata Scans
•
to ensure that partially populated objects can be accessed Column Buffer Cache
both in the IM column store and IM columnar format in Store
Exadata Smart Flash Cache (i.e. cell memory)

29 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Exadata

30 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Hybrid Exadata Scans
• Prior to 23ai, a table scan on Exadata would not
access the In-Memory Column store if less than 80%
of the table was populated In-Memory Columnar scans

• An Exadata Smart Scan was used instead

• The new Hybrid Exadata Scans feature can improve

storage tiering performance with up to 5X gains
• A table scan for a partially populated table will use the In-Memory
Column Store

• The remaining non-populated data can be accessed from Cell Memory In-Flash Columnar scans
Columnar Cache via Exadata Smart Scans

• This can also be a big benefit when using the 23ai

Automatic Sizing feature if the In-Memory column
store size is reduced
• A table scan on Exadata can benefit from end-to-end columnar
processing when if the population percentage of segments fluctuates

31 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Selective In-Memory Columns
New INMEMORY Column Include/Exclude Syntax

• The Wide Tables feature in 23ai allows up to 4096 columns per table, so we have
modified the "NO INMEMORY" syntax to make it easier to manage tables with large
numbers of columns

• Previously you could only exclude columns from population by specifying a column list
in the NO INMEMORY subclause
• Now in 23ai you can specify ALL columns as INMEMORY or NO INMEMORY and then
selectively include or exclude columns from population:
• ALTER TABLE <table name> NO INMEMORY(ALL) INMEMORY(col1, col2);

• ALTER TABLE <table name> INMEMORY(ALL) NO INMEMORY(col1, col2);

32 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Analytics

33 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory Vectorization Framework
Multi-Level Joins and Aggregations

ITEMS
In-Memory Deep Vectorization Framework was
introduced in Oracle Database 21c
• The first feature that leveraged the framework Example:
improved the performance of single level hash Join Sales and Items
joins
Vector Register
SALES
In Oracle Database 23ai the In-Memory Deep

Vector Register
Vectorization Framework adds more join support:
• Multiple Levels of Joins
• Multiple Join Keys
• Semi and Outer Join
• Full Grouping and Aggregation
This translates to faster join performance by utilizing
SIMD optimizations during more types of join
processing

34 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Converged Workloads

35 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory Columnar JSON
JSON documents stored in Database In-Memory (and in Cell Memory on Storage Nodes
with Exadata) automatically get shredded into columns for faster key/value access:

jdoc Path/Value Columns

{
"firstName": ”John", name gender address.city address.state doc id
"gender": ”male",
“age” : 34,
"address": {"city": ”Redwood City", "state": "CA"}, New York
Alan Male NY 2
} Dallas
Clara Female TX 3
{ Redwood
"firstName": ”Alan", John Male CA 1
City
"gender": ”male",
“age” : 24,
"address": {"city": ”New York", "state": ”NY"},
}
SELECT count(*) FROM employee WHERE json_exists
(jdoc, '$.person?(@.age < 34 && @.name = 'John’ &&
{
"firstName": ”Clara", @.address.city = ‘Redwood City’)’)
"gender": ”female",

15X Faster Performance

“age” : 53,
"address": {"city": ”Dallas", "state": ”TX"},
}

3
36 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates
 In-Memory Optimized Dates
Accelerating DATE Queries
Consider a query to find the Total sales amount for every month in 2022
select IM Column Store
Extract(month)
extract(month from order_date) MONTH, order_date
sum(order_amount) TOTAL_SALES Extract(year)
from SALES
where extract(year from order_date) = 2022
group by extract(month from order_date);

• In-Memory can now run such queries up-to 6X faster by

leveraging the In-Memory Expressions framework

• Each extracted component (e.g. MONTH) for a DATE column adds

only a 1B per-row in-memory overhead
• User can specify which DATE column component should be
stored in-memory through a parameter

37 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Autonomous Database

39 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory on Autonomous Database

ADB tenants today automatically leverage Cell-Memory today. But In-Memory can be much faster.

In-Memory Speedups over Cell Memory

8
5.8X 5.3X
6

3.6X 4X
4
2.3X 2X
2

0
SSB SF=300
SSB SF=300 EML
INTERNAL_1 MERCK
INTERNAL_2 DREY
INTERNAL_3 TPC-DS SF=300 TPC-H
TPC-DS SF=300 SF=300
TPC-H SF=300

In-Memory is now supported on Autonomous Database Dedicated!

As of Oracle Autonomous Database Dedicated version 19.20.0.1.0, you can enable Database In-Memory:
You must allocate 4 or more OCPUs.
You can enable between 5 and 70 percent of the available SGA per ADB.

Oracle DatabaseWorld @ CloudWorld Copyright © 2023, Oracle and/or its affiliates

40
 Oracle DatabaseWorld @ CloudWorld Copyright © 2023, Oracle and/or its affiliates
42
 In-Memory on ADW (Shared and Dedicated)
Yes!! It’s finally here!!

43 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Bring the fastest columnar tier to Autonomous DB

In-Memory
(DB DRAM)

Faster Columnar Scan

XRMEM Cell-Memory
(Storage DRAM)

Cell-Memory
(Storage Flash)

HCC Smart Scan

(Disk)

44 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Fusion ERP: In-Memory is great with DMLs
In-Memory ADB-S Limited Availability highlight

Fusion ERP runs daily bulk update job using a single large transaction
Performance of concurrent queries drops by 12x without In-Memory (CR overhead)
In-Memory queries can ignore the active transaction because of the IM Private Journal (non RAC only)

Query elapsed time for Non In-Memory vs. In-Memory

Query Executions
Bulk update Bulk update
46
starts commits
Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates
 Enterprise Performance Management with In-Memory
The Archives database tracks EPM customer metrics
• There are hundreds of partitions/tables – not all segments can fit In-Memory
• Queries typically access the newest partitions
• Daily report job takes 8 hours to complete
60GB of In-Memory space was provisioned based on 15 OCPU ATP shape
• I/O activity disappeared almost immediately after IM was enabled
• Daily report job now takes 4 hours to complete
• Ad hoc queries now complete in 15 seconds vs. more than a minute previously
• Auto IM works because good performance continues as new monthly partitions are created

47 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 High-Level
In-Memory ADB-S General Availability

Customers can mark a table or partition with the INMEMORY clause

Table scans will transparently trigger In-Memory population

• INMEMORY PRIORITY clause can be used to request proactive population without scans

Populate will transparently trigger In-Memory space to grow

When there is space pressure:

• Auto IM will prioritize the hottest objects for population

• Partially populated objects can be scanned from both In-Memory and Cell-Memory

PDB close, CPU reduction, and oversubscription adjustment can shrink In-Memory space

48 Copyright © 2024, Oracle and/or its affiliates

 In-Memory ADB-S General Availability
New Features

23ai IM Area Shrink Space elasticity

23ai IM Hybrid Exadata Scans Graceful tiering

23ai CDB-level Auto IM Prioritize hot data

ADB IM Area Grow from Populates Space elasticity Autonomous

ADB Overprovisioning Adjustment Fairness
Database
Important IM bug fixes! Continuous

Future IM Advisor 2.0

Future IM External Table + AIM, FastStart

Future All objects INMEMORY by default

49 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory ADB-S General Availability
How to enable

Users need to mark objects with INMEMORY clause

In-Memory data could be auto distributed on different instances when CPU count is high

Service MEDIUM or HIGH is required to take full advantage of the distributed In-Memory footprint

Users need to know the available In-Memory space to fit their dataset
• Users can check sga_target value from v$parameter

• By default, IM can use up to 50% of SGA, which is modifiable with DBMS_INMEMORY_ADMIN.SET_SGA_PERCENTAGE

• Users can monitor v$inmemory_area and v$im_segments for population status

50 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 In-Memory ADB-S General Availability
More details (TBD : Clean Up)

IM not available for ADB-S Free

IM not available for less than 16 ECPUs – ADB-D requires a minimum of 4 OCPUs
IM not available for OCPU model
IM license not required – available for both BYOL and License Included

51 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Internal Customer Feedback

“Our Daily Trend reports improved by 5x by enabling In-Memory. Our day-to-day

Enterprise
queries are performing better even as new data is loaded and the full dataset does
Performance not fit In-Memory. Our trend reports used to take more than 24 hours and now
Monitoring they complete in 5 hours”

Fusion Enterprise “Significant boost seen” with IM vs Cell Memory (4x gains) “
Resource Planning

Application
“Queries in some cases that took 56 seconds to complete ran in 400 milliseconds for
Performance the same data set when using In-Memory”
Monitoring

“Power BI Vertipaq and Tableau Hyper are faster than OAC today. We do see a
Analytics Cloud throughput increase with use of IM and we see specific queries which are faster”

52 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

53 Copyright © 2024, Oracle and/or its affiliates
 24 OCPUs on 2 instances

Per instance
• 12 OCPUs
• 20 GB In-Memory

54 Copyright © 2024, Oracle and/or its affiliates

 HCC is default on ADW
IM can be 2x larger than
HCC

Table distributed across

both instances

55 Copyright © 2024, Oracle and/or its affiliates

 Must use MEDIUM or
HIGH to get PQ

PQ is needed if table is
distributed over RAC

IM query: 2s
Non-IM query: 16s

56 Copyright © 2024, Oracle and/or its affiliates

 Serverless challenges
In-Memory on ADB-S

Space must be elastic

• Must not be reserved upfront

• Must be bounded by PDB sga_target CDB
• Must shrink when PDB closes or CPU count changes
PDB PDB PDB PDB
CDB is overprovisioned Bi-weekly
RAC
• Can be overprovisioned by up to 2x, but most PDBs are idle rolling patch
• Space usage must be fair across PDBs during overprovisioning PDB PDB PDB PDB

Instances are volatile

• Bi-weekly rolling instance restart for patching

• PDB can split/merge/reopen on different instances

57 Copyright © 2024, Oracle and/or its affiliates

 General Availability (June 2024)
In-Memory on ADB-S

Space must be elastic

• Any customer can use In-Memory without ops intervention

CDB
• IM populates will automatically grow the In-Memory Area

PDB PDB PDB PDB

CDB is overprovisioned Bi-weekly
RAC
• In-Memory Area size will be continuously adjusted for fairness rolling patch
• CDB and PDB-level Auto IM will be enabled by default PDB PDB PDB PDB

Instances are volatile

• IM Fast Start will be enabled by default

58 Copyright © 2024, Oracle and/or its affiliates

 Customer Use-Cases
Talk about ADW use-cases also.

59 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Hisense

Database In-Memory Customer

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
Hisense
What’s the Use-Case?

Hisense is a global enterprise in household items and consumer electronics.

They use SAP ERP system for both domestic and international business
• Domestic runs on traditional HW and overseas business on Exadata (previously SUN T5 systems)

Month-end processing period encounters resource contention and queuing delays

• Analytic workloads are slow, especially during peak periods when > 1300 users

Significant data growth due to increased business (+200GB/month)

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
Hisense
How did In-Memory help

Ø Complete Application transparency – no modification needed

Ø Compressed 1.5TB on disk data to 133GB in memory (41X in best case)
Ø Average performance of top monthly closing analytic queries improved by up to 3X
• Hisense estimates that over 8000 analyst hours can be saved each month
• With upgrade to Exadata and In-Memory, execution time decreased by 2X, even
with a 16% increase in load

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
Hisense
In-Memory Features and Technologies used:
• Compression Advisor
• In-Memory Advisor (used during evaluation and production)
• IM Duplicate (with Exadata) –fast performance because of local access
• In-Memory Distribute for Service on RAC
• Foundational Features (e.g. SIMD Vectorization, IM Storage Indexes)

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
 Stef-IT

Database In-Memory Customer

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
Stef-IT
What’s the Use-Case?

STEF is a European leader in the food industry supply chain

They provide temperature-controlled logistics and transport services
• Food manufacturers utilize STEF’s operational expertise for transport, warehousing,
orders preparation, packaging, etc.
Business managers need detailed and daily information from STEF-IT to manage their
activity, such as support decision process, warehouse management, and sales
opportunity
Multiple Datamarts maintain with refreshed data – e.g. Transport and Billing
HTAP processing – high transactional workloads mixed with analytic processing
• Daily data-processing can last 6 hours w/ average throughput of 250 DML ops per secs

Star-Schema - Large fact table (> 4B rows)

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
Stef-IT
How did In-Memory help?

Queries improved by 10-100x out of the box.

Business users can generate analytic reports in near real-time within HTAP environment
• 100M updates a day

In-Memory Features and Technologies used:

• In-Memory Join Improvements (bloom filter)
• Fast Repopulation using techniques like Double-Buffering, Incremental and Trickle Repopulation
• Automatic In-Memory Management (hot partitions kept in memory)
• Fast-Start
• Foundational Features (e.g. SIMD Vectorization, IM Storage Indexes)

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
 MyToys

Database In-Memory Customer

Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
myToys
What’s the Use-Case?

Germany’s leading online store for family and children items

Problem/Need:
• Reports about what’s in stock, outgoing goods, incoming
goods, packages went out to customers (incl. returns), price
changes, etc.
• DWH / Reporting on Exadata X7 & Database In-Memory
• DB size 15TB, In-Memory size 250GB
• Analytics on most recent partitions
Outcome:
• Performance gains: 3X – 10X
• Reduced ETL / less CPU with IM Materialized Views and IM
Expressions
• Self Service BI with DBIM & Tableau
Oracle CloudWorld Copyright © 2022, Oracle and/or its affiliates | Confidential: Restricted
 Summary

69 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory Innovations
Auto-Sizing Column Store
Auto-Enabling Features
Hybrid Exadata Scans
Join Groups RAC support Join Group Aware Hash
In-Memory Expressions Join
JSON/OSON support Performance In-Memory Vectorization (Mutli-Level
Massive capacity - In-Memory on External Tables: Hive & HDFS Joins and Agg)
Exadata flash Mem optimized Row store – In-Memory Optimized Dates
Auto population policies Fast Ingest In-Memory Columnar JSON
Fast-Start Base Level Selective In-Memory Columns
Active Data Guard Stability Release Native In-Memory Advisor

2014 2018 2019 2021 2024

12.1 12.2 18C 19C 21C 23ai

Pure In-Memory column format Automatic In-Memory Self Managing In-Memory

Scan & Filter on compressed data In-Memory Dynamic Scans In-Memory Spatial Analytics
Fast joins In-Memory External tables In-Memory Text Analytics
Data pruning via storage indexes In-Memory Optimized Arithmetic Hybrid Scans
SIMD vector processing Memoptimized Rowstore – Fast Vector Joins
In-Memory aggregation
Lookup
 Oracle Database In-Memory runs everywhere
Database In-Memory is an option for Oracle Database Enterprise Edition (EE)

Autonomous Exadata Database Base Database Oracle

Database Service Service Database@Azure
Extreme Performance

Cloud

Cloud@Customer Oracle Exadata Oracle Database Commodity

Appliance Hardware

On-Prem

71 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Oracle is the leader in
Translytical Data Platforms
Translytical Data Platforms enable fast
performance on both transactional and
analytic workloads using in-memory
technology and scale-out architecture
Database In-Memory does exactly this:
• Dual-Format Architecture
• Columnar Data formats and SIMD Vector Processing
• Fully integrated with RAC, PQ, Sharding, etc.

How do we keep this huge lead?

• Analytics and Mixed Workload
• Converged Database
• Autonomous Database

72 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory
Other notable quotes from Larry 10 years ago during In-Memory Announcement

Mixed Workloads:
“There’s actually very little overhead in
maintaining the column store in-memory in
addition to Oracle’s traditional row store”
Automatic In-Memory Management:
“You say how much memory you want to use in
the computer, tell me what partitions or tables to
be in memory, and drop your analytic indexes”
Performance:
“Queries run 100 times faster and updates,
inserts run two, three times faster.”

73 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Where Can You
Get More Information?

74 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Database In-Memory Blog

Learn More Database In-Memory Hands-on-Lab

Database In-Memory Documentation

75 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Guidelines and Tools
Help is available …
These guidelines can help…
Actions DBIM technical briefs to help customers get started …
• Validate that use cases fit
Database In-Memory benefits
• Use the In-Memory Advisor to
verify benefit

• Don't "just try it out", use the

guidelines to avoid re-inventing Quick Start Quick Start Implementation
(one pager) Guide Guidelines
the wheel (see here) (see here) (see here)

• Use the latest Oracle Database

release and the latest RU to
benefit from the latest
innovations and fixes
76 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates
 Boost Analytics Performance with Oracle Database In-Memory
http://bit.ly/golivelabs

Free Tenancy Reservation

No requirement to sign up for the

Cloud!

77 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 blogs.oracle.com/in-memory/dbim-resources

78 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates

 Thank You
Please fill out your evaluations for session

Click to add text

79 Oracle DatabaseWorld @ CloudWorld Copyright © 2024, Oracle and/or its affiliates